404K SEMI-AI Evening Tech Brief 2026-07-09 — AI Infrastructure Bottlenecks, Memory Prices Keep Rising, Software Diverges
目录
Premarket Takeaways
Full AI/Semiconductor Supply Chain
AI Models/Applications and Capex
CSP/Cloud Capex
AI Cloud/Data-Center Operators
GPU/CPU/ASIC
HBM/DRAM/NAND/SSD/HDD
Semiconductor Equipment/Testing
Optical Communications/Optical Chain
High-Speed Interconnect / Power / MLCC
Internet / Platforms
Software / SaaS
Consumer Electronics / Smart Vehicles
Investment-Bank Target Price Changes Over the Past 12 Hours
The AI infrastructure trade is spreading from GPUs alone into memory, optical interconnects, power, equipment, and neo-clouds. Capital is still willing to pay a premium for hard constraints; on the internet and software side, dispersion is emerging, and agent products need to prove revenue and customer data readiness.
Premarket Takeaways
Tonight’s core technology theme remains AI infrastructure. Semiconductors, memory, optical interconnects, and neo-clouds are stronger, while internet and software are under pressure. Market attention has already expanded from Nvidia GPU shipments to memory, power, CPO, advanced packaging, server networking, and cloud-provider cash flow.
Capital is more tolerant of hard assets and more selective toward software narratives. Memory prices, WFE forecasts, Meta’s 1GW data center, Cerebras’s 200MW Europe plan, and IREN’s Microsoft delivery window all reinforce the view that demand is still there. Salesforce’s downgrade shows that agent products still need to be validated by customer budgets and data quality.
Risks are also clearer. Data-center project delays, pressure on cloud-provider free cash flow, valuation pullbacks in optical modules, and overheated trading in Korean semiconductor ADRs could shift the market from a broad rally to a new ranking based on orders, capacity, and profit delivery.
Full AI/Semiconductor Supply Chain
AI Models/Applications and Capex
OpenAI/Anthropic/xAI
1. The model race continues to accelerate. Grok 4.5 is described as aimed at high-capacity coding and agentic AI, priced at USD 2 per million input tokens and USD 6 per million output tokens, below Claude Opus 4.8’s USD 5/USD 25.
2. GPT-6 may reportedly arrive in the next four to six weeks, though the public release cadence remains to be confirmed. The investment implication is that enterprise AI price competition will push down unit inference costs while continuing to lift demand for training and inference compute.
"Grok 4.5 today... Opus level but cheaper; intense release window."
Claude Code
Developer agents are entering the workflow-governance stage. The new/checkupcommand cleans up unused skills, deduplicates project instructions, splits root-level configuration, disables slow hooks, and confirms before making changes. It does not directly provide revenue or order numbers, but it shows that AI coding competition is shifting from “can generate code” to “reduces context waste and configuration friction.” The next things to watch are actual usage frequency and enterprise seat conversion.
Prime Intellect/Lovable
AI application financing remains active. Prime Intellect completed a USD 130 million Series A. Lovable is in talks to raise USD 300 million at a USD 13.2 billion valuation, double its USD 6.6 billion valuation in December; its annualized revenue run rate reached USD 500 million in June. Whether the application layer can sustain valuations will depend on revenue retention, inference costs, and enterprise procurement cycles.
CSP/Cloud Capex
Meta
Meta’s Canadian data center is tonight’s clearest CSP capex signal: the project is roughly CAD 13 billion, with 1GW of capacity, about 3,000 workers during construction, and more than 300 long-term roles. Another estimate says Meta’s compute capacity will rise from 7.5GW at end-2025 to 21.2GW at end-2028, with AI capacity rising from 1.5GW to 13.2GW. This is positive for servers, power, cooling, and networking, but it will also intensify market scrutiny of free cash flow.
"Meta will invest over CAD 13 billion in first Canadian AI Data Center."
Google/Amazon/Meta
Bank of America raised its capacity estimates for the three major cloud providers from about 27GW at end-2025 to 39GW in 2026 and 57GW in 2027, and gave 2027 capex figures of USD 290 billion for Google, USD 185 billion for Meta, and USD 230 billion for AWS. Cloud capex is still being revised upward, but power access, approvals, and funding costs will determine which suppliers actually receive orders.
Amazon/Microsoft/Google/Meta
Large-platform valuations are still tolerant of AI capex. Under BofA’s estimates, 2027 P/E multiples are 21.4x for Amazon, 18.7x for Microsoft, 16.3x for Meta, and 15.7x for Nvidia, versus about 21x forward for the S&P; 500. Platform-stock valuations do not fully reflect growth expectations. The next test is whether cloud revenue can cover capital spending and depreciation pressure.
AI Cloud/Data-Center Operators
Cerebras
Cerebras plans to launch its first European data-center capacity by the end of 2026 and expand in France, Norway, and Finland, targeting total AI infrastructure capacity of 200MW by the end of 2027. Part of the new capacity is expected to support OpenAI workloads. This development links wafer-scale AI computing, local European compute, and low-latency inference. The next things to watch are customer contracts, deployment cadence, and power costs.
"Total AI infrastructure capacity is targeted to reach 200 MW by the end of 2027."
IREN
IREN’s Horizon 1 is expected to be delivered to Microsoft around July 19 and is the first phase of its 200MW Horizon 1-4 plan. Horizon 1 is expected to contribute roughly USD 120-130 million in quarterly revenue, while the full 200MW build corresponds, under different estimates, to about USD 500 million in annualized revenue or a quarterly revenue ramp. Key validation points are Microsoft delivery confirmation, GB300 certification, and the launch cadence for Horizon 2-4.
Nebius
1. The neo-cloud rebound is driven by the demand gap, but the competitive landscape is becoming more complex. Nebius, CoreWeave, and IREN rebounded by roughly +11%, +8%, and +8%, respectively. If Meta rents out excess GPU capacity, a key customer becomes a competitor. The positive is that AI compute demand still exceeds available supply; the risk is that new supply pushes down rental prices and gross margins.
2. Nebius’s operating slope is steep: Q1 2026 revenue was USD 399 million, up +684% YoY; end-2026 ARR guidance is USD 7-9 billion; capex guidance was raised to USD 20-25 billion; and the core AI business has a 45% adjusted EBITDA margin. Behind the high growth, investors need to watch customer concentration, funding costs, and GPU utilization.
GPU/CPU/ASIC
Nvidia
1. Citi maintained Buy and a USD 300 target price, calling Nvidia its preferred large data-center semiconductor name. Management says the Rubin roadmap is fully intact, Spectrum-X scale-out CPO is in production, and customer adoption interest is high.
2. China may approve limited purchases of fewer than 200,000 H200 chips by Alibaba, ByteDance, DeepSeek, and others, corresponding to an estimated USD 4-6 billion in revenue.
3. BofA estimates that by 2030, Nvidia hardware shipments alone will correspond to 99GW of data-center capacity; including ASICs and AMD could push the figure above 150GW.
4. Nvidia rack economics: from Blackwell to Vera Rubin, HBM cost per rack rises by about USD 200,000-300,000, but rack pricing is expected to increase from USD 3-4 million to USD 6-7 million, supporting gross margin in the mid-70% range. The market is worried about HBM costs and ASIC competition. The counter-evidence to watch is rack ASP, long-term HBM contract pricing, and customer deployment cadence.
"roadmap is fully intact; CPO for scale-out is in production with Spectrum-X."
Broadcom
Broadcom has regained certainty from both Apple and AI ASICs. Apple extended its ASIC partnership through 2031 and committed more than USD 30 billion of RF/FBAR chip spending, implying roughly five to six years of USD 5-6 billion in annual revenue. Broadcom’s AI semiconductor revenue last quarter was USD 10.8 billion, up +143% YoY, with guidance for roughly USD 16 billion next quarter. The risk is that Apple’s in-house N1 chip weakens demand for parts of the wireless combo-chip portfolio.
"ASIC deal extended through 2031; AI semis already did 10.8 billion last quarter."
AMD/ASIC Ecosystem
AI infrastructure capacity estimates are now placing AMD and ASICs in the same table. If Nvidia hardware corresponds to 99GW by 2030, and ASIC plus AMD deployments may push total capacity above 150GW, it shows demand is spreading from a single GPU supplier to multiple chip routes. AMD’s validation points shift to MI-series delivery, CoWoS allocation, HBM procurement, and cloud-customer deployments.
Intel/Samsung Electronics AI PC
Edge AI chips are heating up. Samsung Electronics’ GAIA AI PC accelerator uses a 4nm process and has already provided prototypes to Lenovo, HP, and other customers for validation. Mass production could start as early as next year, and Samsung is also trying to combine it with PIM compute-memory technology. On Intel’s side, Nova Lake-S/-H CPUs are expected to include a 12-core Arc Xe3P iGPU. The key for PC AI lies in OEM mass production, software calls, and memory costs.
HBM/DRAM/NAND/SSD/HDD
Memory Industry
Memory is tonight’s clearest pricing line. UBS and Bernstein said global memory sales reached USD 74.6 billion, up +31.7% MoM; DRAM reached USD 48.0 billion, up +27.7% MoM; and NAND reached USD 25.8 billion, up +40.7% MoM. UBS expects DDR contract prices to rise +32% in Q3 2026 and +18% in Q4, while NAND rises +30% in Q3 and +12% in Q4. The upside risk is continued AI-server hoarding; the downside risk is demand destruction on the consumer side.
"Global memory sales reached a record US 74.6 billion, up 31.7% MoM."
Micron/SanDisk/Western Digital/Seagate Technology
The AI storage chain is expanding from HBM to NAND, SSDs, and HDDs. UBS expects HBM demand to rise +90% in 2026 and +77% in 2027, with DRAM structurally undersupplied at least until Q2 2028. The market also cites Micron, SanDisk, Western Digital, and Seagate Technology as beneficiaries of memory price increases. The next things to watch are long-term contract pricing, enterprise SSD shipments, and whether customers accept price increases.
SK Hynix
SK Hynix’s U.S. offering was more than seven times oversubscribed, with market reports saying it involved 177.9 million ADR shares. KB maintained Buy and a KRW 4.2 million target price, and expects 2027 DRAM/NAND capacity growth of only 7%/4%, below demand growth of 17%/19%. Financing enthusiasm supports valuation, but order delivery still depends on HBM customer qualification and the use of capex.
Samsung Electronics/ChangXin Memory Technologies
Samsung Electronics’ GAIA, combined with memory price increases, reinforces its “memory + logic + AI PC” vertical-integration narrative. ChangXin Memory Technologies disclosed its STAR Market listing schedule, with online and offline subscriptions set for July 16. Apple was also reported to have begun testing ChangXin DRAM to increase its negotiating leverage. The impact of domestic DRAM depends on yield, customer qualification, and pricing; a single IPO milestone cannot independently determine the industrial impact.
Phison/Penguin Solutions
Storage control and AI-server support are also translating into results. Phison’s June revenue was TWD 24.853 billion, up +9% MoM and +301% YoY; Q2 revenue was TWD 67.888 billion, up +66% QoQ and +280% YoY; and first-half revenue was TWD 108.855 billion, up +243% YoY. Penguin Solutions had its target price raised by both Stifel and Citizens, citing a higher share of AI-driven business and sequential strength in the memory business.
Semiconductor Equipment/Testing
Applied Materials
1. Deutsche Bank raised its WFE forecasts to USD 145 billion in 2026, USD 193 billion in 2027, and USD 220 billion in 2028, expecting AI-driven fab spending to accelerate through 2028. Applied Materials remains its top pick, with its target price raised from USD 550 to USD 680. Lam Research remains Buy, with its target raised from USD 325 to USD 385. KLA remains Hold, with its target raised from USD 175 to USD 220. Equipment rerating is more biased toward capacity-driven segments, and upside will diverge by subsegment.
2. Separately, Mizuho raised its Applied Materials target price from USD 540 to USD 650 and maintained Buy. Applied Materials’ CEO told Nikkei that chipmakers are preparing for multi-year expansion and are sharing demand outlooks of two years or more with equipment suppliers. Improved order visibility is positive, but the risk is that if memory and foundry capex are pulled forward, the cycle after 2028 will be steeper.
"AI-driven fab spending to accelerate through 2028; WFE forecasts to USD 145 billion in 2026."
AEHR/FormFactor
Silicon photonics and CPO are pushing testing to the front of the supply chain. AEHR received additional WLBI system follow-on orders from SiPho customers and already has PLBI system orders from major hyperscale customers. FormFactor’s TRITON is described as an integrated silicon-photonics test cell covering optical and electrical wafer testing. Photonic-product testing is different from HBM and CPUs. If CPO ramps, test throughput will become a new bottleneck.
Optical Communications/Optical Chain
Lumentum
Lumentum has become a core beneficiary of AI optical networking. Fiscal Q3 2026 revenue was USD 808.4 million, nearly doubling; components revenue was USD 533.3 million, up +77% YoY; systems revenue was USD 275.1 million, up +121% YoY; and current-quarter guidance is USD 960 million to USD 1 billion. Pump lasers grew +80% YoY, and multiple AI products are sold out. Northland raised its target price to USD 1,200 and maintained Outperform. The risk is high valuation, with a rolling P/E of about 128x and a roughly 19% pullback over the past month.
"demand is not the constraint, supply is."
AAOI/Coherent
High-beta optical modules are still being priced around capacity delivery. AAOI was cited for its 800G/1.6T capacity ramp, targeting about USD 471 million in monthly transceiver revenue by mid-2027, with demand expected to exceed capacity through Q3 2027. On the day Lumentum pulled back, AAOI fell 17% and Coherent fell 10%, showing that strong fundamentals do not mean valuations will not move.
Semtech / Optical Interconnect Path
The high-speed interconnect debate is shifting from copper to optics. The industry view is that ACC’s problem is reach: 1.6T is already difficult, and 3.2T is close to infeasible. If Rubin Ultra needs to connect eight Oberon racks, Semtech’s ACC solution may lose viability. The investment implication is that inter-rack interconnects are more likely to favor optical modules, lasers, silicon photonics, CPO, and high-speed switching networks.
"Problem with ACC is reach; hard enough at 1.6T, near impossible at 3.2T."
Nokia
Nokia’s defense AI partnership with NestAI reinforces the edge-networking and sovereign-AI themes. In November 2025, Nokia and Tesi announced a EUR100 million joint investment to develop deployable battlefield capabilities covering 5G, NestOS, wireless-network planning models, early detection, and multi-sensor tracking. This is more about European defense AI networks and does not map to traditional optical-module orders.
High-Speed Interconnect / Power / MLCC
Arista / Credo / High-Speed Interconnect
In market performance, Arista rose 9%; Credo, Astera Labs, and high-speed interconnects all sit in the AI server networking and inter-rack connectivity chain. The core logic is that as compute expands from single racks to cross-rack and cross-cluster deployments, network throughput, SerDes, AEC, CPO, and Ethernet switching all need to keep pace. Watch 800G/1.6T shipments, CPO attach rates, and cloud vendors’ in-house networking roadmaps.
Bloom Energy / Vicor / Power
Power is becoming a constraint for AI infrastructure. Baird maintained a USD310 target price on Bloom Energy, estimating backlog rose from USD6 billion at the end of last year to a conservative level above USD14 billion. Vicor and power modules have also been placed into the data-center power-supply chain. Bloom also said its current supply chain can support up to 25GW of annual capacity. Risks are concentrated in power-order conversion and regulatory approval timelines.
